A Behavior Modeling Framework and Basic Behavior Discovery using Affinity Graphs

by
Xinwei Xue

Advised by
Thomas Henderson

Automated human or animal behavior analysis from video sequences is now a very hot research area, which have wide applications in video surveillance, medicine, psychology, etc. We propose a general framework for the behavior modeling, analysis and synthesis. This allows a better understanding and evaluation of the nature and role of behavior models in the following disciplines: ethology, robot behavior specifications, animated character behavior specification, and automatic behavior analysis. In addition, it is argued that the following aspects a system's behavior usually characterize behavior models: (1) physical, (2), physiological, (3) conceptual, and (4) contextual. We also demonstrate how basic behavior units can be extracted from time sequence data of physical systems, animals, and humans. This is done by constructing an affinity graph from temporal sequences that are similar as defined in a correlation measure, and then finding the eigenvectors of the affinity graph in order to cluster similar behavior sequences.